Comment by viraptor
5 months ago
It seems to be slowing down actually. Last year was wild until around llama 3. The latest improvements are relatively small. Even the reasoning models are a small improvement over explicit planning with agents that we could already do before - it's just nicely wrapped and slightly tuned for that purpose. Deepseek did some serious efficiency improvements, but not so much user-visible things.
So I'd say that the AI race is starting to plateau a bit recently.
While I agree, you have to remember the dimensionality of the labor-skill space is. The was I see it is that you can imagine the capability of AI as a radius, and the amount of tasks it can cover is a sphere. Linear imporovements in performance causes cubic (or whatever the labor-skill dimensionality is) imporvement in task coverage.
I‘m not sure that’s true with the latest models. o3-mini is good at analytical tasks and coding, and it really sucks at prose. Sonnet 3.7 is good at thinking but lost some ability in creating diffs.